numam-spdk/lib/env_dpdk/memory.c

1454 lines
38 KiB
C
Raw Normal View History

/*-
* BSD LICENSE
*
* Copyright (c) Intel Corporation.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
* * Neither the name of Intel Corporation nor the names of its
* contributors may be used to endorse or promote products derived
* from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "spdk/stdinc.h"
#include "env_internal.h"
#include <rte_config.h>
env: Use rte_malloc in spdk_mem_register code path when possible For TCMalloc regions which we register with spdk at runtime in the MMapHook, we need to ensure that SPDK doesn't do any allocations in that path otherwise we will hit a livelock situation. MmapHook is invoked when TCMalloc is out of free memory and needs to get more memory from the system, for the hugepage case it gets via mmap. In the current code, we could end up calling malloc in the spdk_mem_register call via the following call path. spdk_mem_register -> spdk_mem_map_set_translation -> spdk_mem_map_get_map_1gb To avoid this livelock situation we call rte_malloc instead which shouldn't invoke the system allocator. Note that in try_expand_heap_primary() which is invoked in the rte_malloc code path, we can still call malloc, so we need to only use this when dynamic memory allocation is disabled via --legacy-mem. It is possible in the future we could work around even this limitation, but for now this implementation will be much simpler. Have verified this change fixes the livelock condition which I was hitting in my setup without this fix. Change-Id: I69d0813a70da1f26f8c4d9d8895e406c026be18b Signed-off-by: Alok N Kataria <alok.kataria@nutanix.com> Signed-off-by: Jim Harris <james.r.harris@intel.com> Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/475943 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Community-CI: SPDK CI Jenkins <sys_sgci@intel.com>
2019-11-27 05:00:17 +00:00
#include <rte_malloc.h>
#include <rte_memory.h>
#include <rte_eal_memconfig.h>
#include "spdk_internal/assert.h"
#include "spdk/assert.h"
#include "spdk/likely.h"
#include "spdk/queue.h"
#include "spdk/util.h"
#include "spdk/memory.h"
#include "spdk/env_dpdk.h"
#ifdef __FreeBSD__
#define SPDK_VFIO_ENABLED 0
#else
#include <linux/version.h>
#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 6, 0)
#define SPDK_VFIO_ENABLED 1
#include <linux/vfio.h>
#include <rte_vfio.h>
struct spdk_vfio_dma_map {
struct vfio_iommu_type1_dma_map map;
struct vfio_iommu_type1_dma_unmap unmap;
TAILQ_ENTRY(spdk_vfio_dma_map) tailq;
};
struct vfio_cfg {
int fd;
bool enabled;
bool noiommu_enabled;
unsigned device_ref;
TAILQ_HEAD(, spdk_vfio_dma_map) maps;
pthread_mutex_t mutex;
};
static struct vfio_cfg g_vfio = {
.fd = -1,
.enabled = false,
.noiommu_enabled = false,
.device_ref = 0,
.maps = TAILQ_HEAD_INITIALIZER(g_vfio.maps),
.mutex = PTHREAD_MUTEX_INITIALIZER
};
#else
#define SPDK_VFIO_ENABLED 0
#endif
#endif
#if DEBUG
#define DEBUG_PRINT(...) fprintf(stderr, __VA_ARGS__)
#else
#define DEBUG_PRINT(...)
#endif
#define FN_2MB_TO_4KB(fn) (fn << (SHIFT_2MB - SHIFT_4KB))
#define FN_4KB_TO_2MB(fn) (fn >> (SHIFT_2MB - SHIFT_4KB))
#define MAP_256TB_IDX(vfn_2mb) ((vfn_2mb) >> (SHIFT_1GB - SHIFT_2MB))
#define MAP_1GB_IDX(vfn_2mb) ((vfn_2mb) & ((1ULL << (SHIFT_1GB - SHIFT_2MB)) - 1))
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
/* Page is registered */
#define REG_MAP_REGISTERED (1ULL << 62)
/* A notification region barrier. The 2MB translation entry that's marked
* with this flag must be unregistered separately. This allows contiguous
* regions to be unregistered in the same chunks they were registered.
*/
#define REG_MAP_NOTIFY_START (1ULL << 63)
/* Translation of a single 2MB page. */
struct map_2mb {
uint64_t translation_2mb;
};
/* Second-level map table indexed by bits [21..29] of the virtual address.
* Each entry contains the address translation or error for entries that haven't
* been retrieved yet.
*/
struct map_1gb {
struct map_2mb map[1ULL << (SHIFT_1GB - SHIFT_2MB)];
};
/* Top-level map table indexed by bits [30..47] of the virtual address.
* Each entry points to a second-level map table or NULL.
*/
struct map_256tb {
struct map_1gb *map[1ULL << (SHIFT_256TB - SHIFT_1GB)];
};
/* Page-granularity memory address translation */
struct spdk_mem_map {
struct map_256tb map_256tb;
pthread_mutex_t mutex;
uint64_t default_translation;
struct spdk_mem_map_ops ops;
void *cb_ctx;
TAILQ_ENTRY(spdk_mem_map) tailq;
};
/* Registrations map. The 64 bit translations are bit fields with the
* following layout (starting with the low bits):
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
* 0 - 61 : reserved
* 62 - 63 : flags
*/
static struct spdk_mem_map *g_mem_reg_map;
static TAILQ_HEAD(spdk_mem_map_head, spdk_mem_map) g_spdk_mem_maps =
TAILQ_HEAD_INITIALIZER(g_spdk_mem_maps);
static pthread_mutex_t g_spdk_mem_map_mutex = PTHREAD_MUTEX_INITIALIZER;
static bool g_legacy_mem;
/*
* Walk the currently registered memory via the main memory registration map
* and call the new map's notify callback for each virtually contiguous region.
*/
static int
spdk_mem_map_notify_walk(struct spdk_mem_map *map, enum spdk_mem_map_notify_action action)
{
size_t idx_256tb;
uint64_t idx_1gb;
uint64_t contig_start = UINT64_MAX;
uint64_t contig_end = UINT64_MAX;
struct map_1gb *map_1gb;
int rc;
if (!g_mem_reg_map) {
return -EINVAL;
}
/* Hold the memory registration map mutex so no new registrations can be added while we are looping. */
pthread_mutex_lock(&g_mem_reg_map->mutex);
for (idx_256tb = 0;
idx_256tb < sizeof(g_mem_reg_map->map_256tb.map) / sizeof(g_mem_reg_map->map_256tb.map[0]);
idx_256tb++) {
map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
if (!map_1gb) {
if (contig_start != UINT64_MAX) {
/* End of of a virtually contiguous range */
rc = map->ops.notify_cb(map->cb_ctx, map, action,
(void *)contig_start,
contig_end - contig_start + VALUE_2MB);
/* Don't bother handling unregister failures. It can't be any worse */
if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
goto err_unregister;
}
}
contig_start = UINT64_MAX;
continue;
}
for (idx_1gb = 0; idx_1gb < sizeof(map_1gb->map) / sizeof(map_1gb->map[0]); idx_1gb++) {
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
if ((map_1gb->map[idx_1gb].translation_2mb & REG_MAP_REGISTERED) &&
(contig_start == UINT64_MAX ||
(map_1gb->map[idx_1gb].translation_2mb & REG_MAP_NOTIFY_START) == 0)) {
/* Rebuild the virtual address from the indexes */
uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << SHIFT_2MB);
if (contig_start == UINT64_MAX) {
contig_start = vaddr;
}
contig_end = vaddr;
} else {
if (contig_start != UINT64_MAX) {
/* End of of a virtually contiguous range */
rc = map->ops.notify_cb(map->cb_ctx, map, action,
(void *)contig_start,
contig_end - contig_start + VALUE_2MB);
/* Don't bother handling unregister failures. It can't be any worse */
if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
goto err_unregister;
}
/* This page might be a part of a neighbour region, so process
* it again. The idx_1gb will be incremented immediately.
*/
idx_1gb--;
}
contig_start = UINT64_MAX;
}
}
}
pthread_mutex_unlock(&g_mem_reg_map->mutex);
return 0;
err_unregister:
/* Unwind to the first empty translation so we don't unregister
* a region that just failed to register.
*/
idx_256tb = MAP_256TB_IDX((contig_start >> SHIFT_2MB) - 1);
idx_1gb = MAP_1GB_IDX((contig_start >> SHIFT_2MB) - 1);
contig_start = UINT64_MAX;
contig_end = UINT64_MAX;
/* Unregister any memory we managed to register before the failure */
for (; idx_256tb < SIZE_MAX; idx_256tb--) {
map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
if (!map_1gb) {
if (contig_end != UINT64_MAX) {
/* End of of a virtually contiguous range */
map->ops.notify_cb(map->cb_ctx, map,
SPDK_MEM_MAP_NOTIFY_UNREGISTER,
(void *)contig_start,
contig_end - contig_start + VALUE_2MB);
}
contig_end = UINT64_MAX;
continue;
}
for (; idx_1gb < UINT64_MAX; idx_1gb--) {
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
if ((map_1gb->map[idx_1gb].translation_2mb & REG_MAP_REGISTERED) &&
(contig_end == UINT64_MAX || (map_1gb->map[idx_1gb].translation_2mb & REG_MAP_NOTIFY_START) == 0)) {
/* Rebuild the virtual address from the indexes */
uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << SHIFT_2MB);
if (contig_end == UINT64_MAX) {
contig_end = vaddr;
}
contig_start = vaddr;
} else {
if (contig_end != UINT64_MAX) {
/* End of of a virtually contiguous range */
map->ops.notify_cb(map->cb_ctx, map,
SPDK_MEM_MAP_NOTIFY_UNREGISTER,
(void *)contig_start,
contig_end - contig_start + VALUE_2MB);
idx_1gb++;
}
contig_end = UINT64_MAX;
}
}
idx_1gb = sizeof(map_1gb->map) / sizeof(map_1gb->map[0]) - 1;
}
pthread_mutex_unlock(&g_mem_reg_map->mutex);
return rc;
}
struct spdk_mem_map *
spdk_mem_map_alloc(uint64_t default_translation, const struct spdk_mem_map_ops *ops, void *cb_ctx)
{
struct spdk_mem_map *map;
int rc;
map = calloc(1, sizeof(*map));
if (map == NULL) {
return NULL;
}
if (pthread_mutex_init(&map->mutex, NULL)) {
free(map);
return NULL;
}
map->default_translation = default_translation;
map->cb_ctx = cb_ctx;
if (ops) {
map->ops = *ops;
}
if (ops && ops->notify_cb) {
pthread_mutex_lock(&g_spdk_mem_map_mutex);
rc = spdk_mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_REGISTER);
if (rc != 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
DEBUG_PRINT("Initial mem_map notify failed\n");
pthread_mutex_destroy(&map->mutex);
free(map);
return NULL;
}
TAILQ_INSERT_TAIL(&g_spdk_mem_maps, map, tailq);
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
}
return map;
}
void
spdk_mem_map_free(struct spdk_mem_map **pmap)
{
struct spdk_mem_map *map;
size_t i;
if (!pmap) {
return;
}
map = *pmap;
if (!map) {
return;
}
if (map->ops.notify_cb) {
pthread_mutex_lock(&g_spdk_mem_map_mutex);
spdk_mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_UNREGISTER);
TAILQ_REMOVE(&g_spdk_mem_maps, map, tailq);
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
}
for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
env: Use rte_malloc in spdk_mem_register code path when possible For TCMalloc regions which we register with spdk at runtime in the MMapHook, we need to ensure that SPDK doesn't do any allocations in that path otherwise we will hit a livelock situation. MmapHook is invoked when TCMalloc is out of free memory and needs to get more memory from the system, for the hugepage case it gets via mmap. In the current code, we could end up calling malloc in the spdk_mem_register call via the following call path. spdk_mem_register -> spdk_mem_map_set_translation -> spdk_mem_map_get_map_1gb To avoid this livelock situation we call rte_malloc instead which shouldn't invoke the system allocator. Note that in try_expand_heap_primary() which is invoked in the rte_malloc code path, we can still call malloc, so we need to only use this when dynamic memory allocation is disabled via --legacy-mem. It is possible in the future we could work around even this limitation, but for now this implementation will be much simpler. Have verified this change fixes the livelock condition which I was hitting in my setup without this fix. Change-Id: I69d0813a70da1f26f8c4d9d8895e406c026be18b Signed-off-by: Alok N Kataria <alok.kataria@nutanix.com> Signed-off-by: Jim Harris <james.r.harris@intel.com> Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/475943 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Community-CI: SPDK CI Jenkins <sys_sgci@intel.com>
2019-11-27 05:00:17 +00:00
if (g_legacy_mem) {
rte_free(map->map_256tb.map[i]);
} else {
free(map->map_256tb.map[i]);
}
}
pthread_mutex_destroy(&map->mutex);
free(map);
*pmap = NULL;
}
int
spdk_mem_register(void *vaddr, size_t len)
{
struct spdk_mem_map *map;
int rc;
void *seg_vaddr;
size_t seg_len;
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
uint64_t reg;
if ((uintptr_t)vaddr & ~MASK_256TB) {
DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
return -EINVAL;
}
if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
DEBUG_PRINT("invalid %s parameters, vaddr=%p len=%ju\n",
__func__, vaddr, len);
return -EINVAL;
}
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
if (len == 0) {
return 0;
}
pthread_mutex_lock(&g_spdk_mem_map_mutex);
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
seg_vaddr = vaddr;
seg_len = len;
while (seg_len > 0) {
reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
if (reg & REG_MAP_REGISTERED) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return -EBUSY;
}
seg_vaddr += VALUE_2MB;
seg_len -= VALUE_2MB;
}
seg_vaddr = vaddr;
seg_len = 0;
while (len > 0) {
spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, VALUE_2MB,
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
seg_len == 0 ? REG_MAP_REGISTERED | REG_MAP_NOTIFY_START : REG_MAP_REGISTERED);
seg_len += VALUE_2MB;
vaddr += VALUE_2MB;
len -= VALUE_2MB;
}
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_REGISTER, seg_vaddr, seg_len);
if (rc != 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return rc;
}
}
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return 0;
}
int
spdk_mem_unregister(void *vaddr, size_t len)
{
struct spdk_mem_map *map;
int rc;
void *seg_vaddr;
size_t seg_len;
uint64_t reg, newreg;
if ((uintptr_t)vaddr & ~MASK_256TB) {
DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
return -EINVAL;
}
if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
DEBUG_PRINT("invalid %s parameters, vaddr=%p len=%ju\n",
__func__, vaddr, len);
return -EINVAL;
}
pthread_mutex_lock(&g_spdk_mem_map_mutex);
/* The first page must be a start of a region. Also check if it's
* registered to make sure we don't return -ERANGE for non-registered
* regions.
*/
reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
if ((reg & REG_MAP_REGISTERED) && (reg & REG_MAP_NOTIFY_START) == 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return -ERANGE;
}
seg_vaddr = vaddr;
seg_len = len;
while (seg_len > 0) {
reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
if ((reg & REG_MAP_REGISTERED) == 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return -EINVAL;
}
seg_vaddr += VALUE_2MB;
seg_len -= VALUE_2MB;
}
newreg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
/* If the next page is registered, it must be a start of a region as well,
* otherwise we'd be unregistering only a part of a region.
*/
if ((newreg & REG_MAP_NOTIFY_START) == 0 && (newreg & REG_MAP_REGISTERED)) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return -ERANGE;
}
seg_vaddr = vaddr;
seg_len = 0;
while (len > 0) {
reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, VALUE_2MB, 0);
if (seg_len > 0 && (reg & REG_MAP_NOTIFY_START)) {
TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
memory: forbid registering a memory region more than once Removed the reference count from the registrations map. Although technically supported, registering a single memory region more than once had a lot of unhandled cases and could easily lead to a segfault. RDMA maps require all memory to be unregistered in the same chunks the memory was registered, which is often impossible to achieve if a region was registered more than once: 1. register region 0x0 - 0x3 -> it gets mapped to a single ibv_mr 2. register region 0x1 - 0x2 -> nothing happens, this region is already registered 3. unregister region 0x0 - 0x3 -> 0x0-0x1 gets unregistered as one region. 0x2-0x3 gets unregistered as another (leading to segfault in the the current RDMA implementation) The problem is that the last two regions share the same ibv_mr, which SPDK tries to free twice. The second free causes a segfault. vtophys map handles this case by registering each 2MB chunk separately, but this solution cannot be applied for RDMA, as NICs put a limitation (~2048) on the number of regions registered. Another option is to keep a refcount of each ibv_mr allocated, and free it only when the entire region was unregistered from the SPDK mem map. This is however very tricky and RDMAmojo mentions that freeing a memory buffer before unregistering its ibv_mr may lead to a segfault. Change-Id: I545c56e24ffa55bda211dea22aeb8a55d9631fe5 Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com> Reviewed-on: https://review.gerrithub.io/426085 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Reviewed-by: Jim Harris <james.r.harris@intel.com> Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
2018-09-19 11:36:22 +00:00
rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER, seg_vaddr, seg_len);
if (rc != 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return rc;
}
}
seg_vaddr = vaddr;
seg_len = VALUE_2MB;
} else {
seg_len += VALUE_2MB;
}
vaddr += VALUE_2MB;
len -= VALUE_2MB;
}
if (seg_len > 0) {
TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER, seg_vaddr, seg_len);
if (rc != 0) {
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return rc;
}
}
}
pthread_mutex_unlock(&g_spdk_mem_map_mutex);
return 0;
}
static struct map_1gb *
spdk_mem_map_get_map_1gb(struct spdk_mem_map *map, uint64_t vfn_2mb)
{
struct map_1gb *map_1gb;
uint64_t idx_256tb = MAP_256TB_IDX(vfn_2mb);
size_t i;
if (spdk_unlikely(idx_256tb >= SPDK_COUNTOF(map->map_256tb.map))) {
return NULL;
}
map_1gb = map->map_256tb.map[idx_256tb];
if (!map_1gb) {
pthread_mutex_lock(&map->mutex);
/* Recheck to make sure nobody else got the mutex first. */
map_1gb = map->map_256tb.map[idx_256tb];
if (!map_1gb) {
env: Use rte_malloc in spdk_mem_register code path when possible For TCMalloc regions which we register with spdk at runtime in the MMapHook, we need to ensure that SPDK doesn't do any allocations in that path otherwise we will hit a livelock situation. MmapHook is invoked when TCMalloc is out of free memory and needs to get more memory from the system, for the hugepage case it gets via mmap. In the current code, we could end up calling malloc in the spdk_mem_register call via the following call path. spdk_mem_register -> spdk_mem_map_set_translation -> spdk_mem_map_get_map_1gb To avoid this livelock situation we call rte_malloc instead which shouldn't invoke the system allocator. Note that in try_expand_heap_primary() which is invoked in the rte_malloc code path, we can still call malloc, so we need to only use this when dynamic memory allocation is disabled via --legacy-mem. It is possible in the future we could work around even this limitation, but for now this implementation will be much simpler. Have verified this change fixes the livelock condition which I was hitting in my setup without this fix. Change-Id: I69d0813a70da1f26f8c4d9d8895e406c026be18b Signed-off-by: Alok N Kataria <alok.kataria@nutanix.com> Signed-off-by: Jim Harris <james.r.harris@intel.com> Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/475943 Tested-by: SPDK CI Jenkins <sys_sgci@intel.com> Reviewed-by: Changpeng Liu <changpeng.liu@intel.com> Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com> Reviewed-by: Ben Walker <benjamin.walker@intel.com> Community-CI: SPDK CI Jenkins <sys_sgci@intel.com>
2019-11-27 05:00:17 +00:00
/* Some of the existing apps use TCMalloc hugepage
* allocator and register this tcmalloc allocated
* hugepage memory with SPDK in the mmap hook. Since
* this function is called in the spdk_mem_register
* code path we can't do a malloc here otherwise that
* would cause a livelock. So we use the dpdk provided
* allocator instead, which avoids this cyclic
* dependency. Note this is only guaranteed to work when
* DPDK dynamic memory allocation is disabled (--legacy-mem),
* which then is a requirement for anyone using TCMalloc in
* this way.
*/
if (g_legacy_mem) {
map_1gb = rte_malloc(NULL, sizeof(struct map_1gb), 0);
} else {
map_1gb = malloc(sizeof(struct map_1gb));
}
if (map_1gb) {
/* initialize all entries to default translation */
for (i = 0; i < SPDK_COUNTOF(map_1gb->map); i++) {
map_1gb->map[i].translation_2mb = map->default_translation;
}
map->map_256tb.map[idx_256tb] = map_1gb;
}
}
pthread_mutex_unlock(&map->mutex);
if (!map_1gb) {
DEBUG_PRINT("allocation failed\n");
return NULL;
}
}
return map_1gb;
}
int
spdk_mem_map_set_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size,
uint64_t translation)
{
uint64_t vfn_2mb;
struct map_1gb *map_1gb;
uint64_t idx_1gb;
struct map_2mb *map_2mb;
if ((uintptr_t)vaddr & ~MASK_256TB) {
DEBUG_PRINT("invalid usermode virtual address %lu\n", vaddr);
return -EINVAL;
}
/* For now, only 2 MB-aligned registrations are supported */
if (((uintptr_t)vaddr & MASK_2MB) || (size & MASK_2MB)) {
DEBUG_PRINT("invalid %s parameters, vaddr=%lu len=%ju\n",
__func__, vaddr, size);
return -EINVAL;
}
vfn_2mb = vaddr >> SHIFT_2MB;
while (size) {
map_1gb = spdk_mem_map_get_map_1gb(map, vfn_2mb);
if (!map_1gb) {
DEBUG_PRINT("could not get %p map\n", (void *)vaddr);
return -ENOMEM;
}
idx_1gb = MAP_1GB_IDX(vfn_2mb);
map_2mb = &map_1gb->map[idx_1gb];
map_2mb->translation_2mb = translation;
size -= VALUE_2MB;
vfn_2mb++;
}
return 0;
}
int
spdk_mem_map_clear_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size)
{
return spdk_mem_map_set_translation(map, vaddr, size, map->default_translation);
}
inline uint64_t
spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
{
const struct map_1gb *map_1gb;
const struct map_2mb *map_2mb;
uint64_t idx_256tb;
uint64_t idx_1gb;
uint64_t vfn_2mb;
uint64_t cur_size;
uint64_t prev_translation;
uint64_t orig_translation;
if (spdk_unlikely(vaddr & ~MASK_256TB)) {
DEBUG_PRINT("invalid usermode virtual address %p\n", (void *)vaddr);
return map->default_translation;
}
vfn_2mb = vaddr >> SHIFT_2MB;
idx_256tb = MAP_256TB_IDX(vfn_2mb);
idx_1gb = MAP_1GB_IDX(vfn_2mb);
map_1gb = map->map_256tb.map[idx_256tb];
if (spdk_unlikely(!map_1gb)) {
return map->default_translation;
}
cur_size = VALUE_2MB - _2MB_OFFSET(vaddr);
map_2mb = &map_1gb->map[idx_1gb];
if (size == NULL || map->ops.are_contiguous == NULL ||
map_2mb->translation_2mb == map->default_translation) {
if (size != NULL) {
*size = spdk_min(*size, cur_size);
}
return map_2mb->translation_2mb;
}
orig_translation = map_2mb->translation_2mb;
prev_translation = orig_translation;
while (cur_size < *size) {
vfn_2mb++;
idx_256tb = MAP_256TB_IDX(vfn_2mb);
idx_1gb = MAP_1GB_IDX(vfn_2mb);
map_1gb = map->map_256tb.map[idx_256tb];
if (spdk_unlikely(!map_1gb)) {
break;
}
map_2mb = &map_1gb->map[idx_1gb];
if (!map->ops.are_contiguous(prev_translation, map_2mb->translation_2mb)) {
break;
}
cur_size += VALUE_2MB;
prev_translation = map_2mb->translation_2mb;
}
*size = spdk_min(*size, cur_size);
return orig_translation;
}
#if RTE_VERSION >= RTE_VERSION_NUM(18, 05, 0, 0)
static void
memory_hotplug_cb(enum rte_mem_event event_type,
const void *addr, size_t len, void *arg)
{
if (event_type == RTE_MEM_EVENT_ALLOC) {
spdk_mem_register((void *)addr, len);
#if RTE_VERSION >= RTE_VERSION_NUM(19, 02, 0, 0)
if (!spdk_env_dpdk_external_init()) {
return;
}
#endif
/* Prior to DPDK 19.02, we have to worry about DPDK
* freeing memory in different units than it was allocated.
* That doesn't work with things like RDMA MRs. So for
* those versions of DPDK, mark each segment so that DPDK
* won't later free it. That ensures we don't have to deal
* with that scenario.
*
* DPDK 19.02 added the --match-allocations RTE flag to
* avoid this condition.
*
* Note: if the user initialized DPDK separately, we can't
* be sure that --match-allocations was specified, so need
* to still mark the segments so they aren't freed.
*/
while (len > 0) {
struct rte_memseg *seg;
seg = rte_mem_virt2memseg(addr, NULL);
assert(seg != NULL);
seg->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
addr = (void *)((uintptr_t)addr + seg->hugepage_sz);
len -= seg->hugepage_sz;
}
} else if (event_type == RTE_MEM_EVENT_FREE) {
spdk_mem_unregister((void *)addr, len);
}
}
static int
memory_iter_cb(const struct rte_memseg_list *msl,
const struct rte_memseg *ms, size_t len, void *arg)
{
return spdk_mem_register(ms->addr, len);
}
#endif
int
spdk_mem_map_init(bool legacy_mem)
{
g_legacy_mem = legacy_mem;
g_mem_reg_map = spdk_mem_map_alloc(0, NULL, NULL);
if (g_mem_reg_map == NULL) {
DEBUG_PRINT("memory registration map allocation failed\n");
return -ENOMEM;
}
/*
* Walk all DPDK memory segments and register them
* with the master memory map
*/
#if RTE_VERSION >= RTE_VERSION_NUM(18, 05, 0, 0)
rte_mem_event_callback_register("spdk", memory_hotplug_cb, NULL);
rte_memseg_contig_walk(memory_iter_cb, NULL);
#else
struct rte_mem_config *mcfg;
size_t seg_idx;
mcfg = rte_eal_get_configuration()->mem_config;
for (seg_idx = 0; seg_idx < RTE_MAX_MEMSEG; seg_idx++) {
struct rte_memseg *seg = &mcfg->memseg[seg_idx];
if (seg->addr == NULL) {
break;
}
spdk_mem_register(seg->addr, seg->len);
}
#endif
return 0;
}
bool
spdk_iommu_is_enabled(void)
{
#if SPDK_VFIO_ENABLED
return g_vfio.enabled && !g_vfio.noiommu_enabled;
#else
return false;
#endif
}
struct spdk_vtophys_pci_device {
struct rte_pci_device *pci_device;
TAILQ_ENTRY(spdk_vtophys_pci_device) tailq;
};
static pthread_mutex_t g_vtophys_pci_devices_mutex = PTHREAD_MUTEX_INITIALIZER;
static TAILQ_HEAD(, spdk_vtophys_pci_device) g_vtophys_pci_devices =
TAILQ_HEAD_INITIALIZER(g_vtophys_pci_devices);
static struct spdk_mem_map *g_vtophys_map;
static struct spdk_mem_map *g_phys_ref_map;
#if SPDK_VFIO_ENABLED
static int
vtophys_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
{
struct spdk_vfio_dma_map *dma_map;
uint64_t refcount;
int ret;
refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
assert(refcount < UINT64_MAX);
if (refcount > 0) {
spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
return 0;
}
dma_map = calloc(1, sizeof(*dma_map));
if (dma_map == NULL) {
return -ENOMEM;
}
dma_map->map.argsz = sizeof(dma_map->map);
dma_map->map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
dma_map->map.vaddr = vaddr;
dma_map->map.iova = iova;
dma_map->map.size = size;
dma_map->unmap.argsz = sizeof(dma_map->unmap);
dma_map->unmap.flags = 0;
dma_map->unmap.iova = iova;
dma_map->unmap.size = size;
pthread_mutex_lock(&g_vfio.mutex);
if (g_vfio.device_ref == 0) {
/* VFIO requires at least one device (IOMMU group) to be added to
* a VFIO container before it is possible to perform any IOMMU
* operations on that container. This memory will be mapped once
* the first device (IOMMU group) is hotplugged.
*
* Since the vfio container is managed internally by DPDK, it is
* also possible that some device is already in that container, but
* it's not managed by SPDK - e.g. an NIC attached internally
* inside DPDK. We could map the memory straight away in such
* scenario, but there's no need to do it. DPDK devices clearly
* don't need our mappings and hence we defer the mapping
* unconditionally until the first SPDK-managed device is
* hotplugged.
*/
goto out_insert;
}
ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
if (ret) {
DEBUG_PRINT("Cannot set up DMA mapping, error %d\n", errno);
pthread_mutex_unlock(&g_vfio.mutex);
free(dma_map);
return ret;
}
out_insert:
TAILQ_INSERT_TAIL(&g_vfio.maps, dma_map, tailq);
pthread_mutex_unlock(&g_vfio.mutex);
spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
return 0;
}
static int
vtophys_iommu_unmap_dma(uint64_t iova, uint64_t size)
{
struct spdk_vfio_dma_map *dma_map;
uint64_t refcount;
int ret;
pthread_mutex_lock(&g_vfio.mutex);
TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
if (dma_map->map.iova == iova) {
break;
}
}
if (dma_map == NULL) {
DEBUG_PRINT("Cannot clear DMA mapping for IOVA %"PRIx64" - it's not mapped\n", iova);
pthread_mutex_unlock(&g_vfio.mutex);
return -ENXIO;
}
refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
assert(refcount < UINT64_MAX);
if (refcount > 0) {
spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount - 1);
}
/* We still have outstanding references, don't clear it. */
if (refcount > 1) {
pthread_mutex_unlock(&g_vfio.mutex);
return 0;
}
/** don't support partial or multiple-page unmap for now */
assert(dma_map->map.size == size);
if (g_vfio.device_ref == 0) {
/* Memory is not mapped anymore, just remove it's references */
goto out_remove;
}
ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &dma_map->unmap);
if (ret) {
DEBUG_PRINT("Cannot clear DMA mapping, error %d\n", errno);
pthread_mutex_unlock(&g_vfio.mutex);
return ret;
}
out_remove:
TAILQ_REMOVE(&g_vfio.maps, dma_map, tailq);
pthread_mutex_unlock(&g_vfio.mutex);
free(dma_map);
return 0;
}
#endif
static uint64_t
vtophys_get_paddr_memseg(uint64_t vaddr)
{
uintptr_t paddr;
struct rte_memseg *seg;
#if RTE_VERSION >= RTE_VERSION_NUM(18, 05, 0, 0)
seg = rte_mem_virt2memseg((void *)(uintptr_t)vaddr, NULL);
if (seg != NULL) {
paddr = seg->phys_addr;
if (paddr == RTE_BAD_IOVA) {
return SPDK_VTOPHYS_ERROR;
}
paddr += (vaddr - (uintptr_t)seg->addr);
return paddr;
}
#else
struct rte_mem_config *mcfg;
uint32_t seg_idx;
mcfg = rte_eal_get_configuration()->mem_config;
for (seg_idx = 0; seg_idx < RTE_MAX_MEMSEG; seg_idx++) {
seg = &mcfg->memseg[seg_idx];
if (seg->addr == NULL) {
break;
}
if (vaddr >= (uintptr_t)seg->addr &&
vaddr < ((uintptr_t)seg->addr + seg->len)) {
paddr = seg->phys_addr;
if (paddr == RTE_BAD_IOVA) {
return SPDK_VTOPHYS_ERROR;
}
paddr += (vaddr - (uintptr_t)seg->addr);
return paddr;
}
}
#endif
return SPDK_VTOPHYS_ERROR;
}
/* Try to get the paddr from /proc/self/pagemap */
static uint64_t
vtophys_get_paddr_pagemap(uint64_t vaddr)
{
uintptr_t paddr;
/* Silence static analyzers */
assert(vaddr != 0);
paddr = rte_mem_virt2iova((void *)vaddr);
if (paddr == RTE_BAD_IOVA) {
/*
* The vaddr may be valid but doesn't have a backing page
* assigned yet. Touch the page to ensure a backing page
* gets assigned, then try to translate again.
*/
rte_atomic64_read((rte_atomic64_t *)vaddr);
paddr = rte_mem_virt2iova((void *)vaddr);
}
if (paddr == RTE_BAD_IOVA) {
/* Unable to get to the physical address. */
return SPDK_VTOPHYS_ERROR;
}
return paddr;
}
/* Try to get the paddr from pci devices */
static uint64_t
vtophys_get_paddr_pci(uint64_t vaddr)
{
struct spdk_vtophys_pci_device *vtophys_dev;
uintptr_t paddr;
struct rte_pci_device *dev;
struct rte_mem_resource *res;
unsigned r;
pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
dev = vtophys_dev->pci_device;
for (r = 0; r < PCI_MAX_RESOURCE; r++) {
res = &dev->mem_resource[r];
if (res->phys_addr && vaddr >= (uint64_t)res->addr &&
vaddr < (uint64_t)res->addr + res->len) {
paddr = res->phys_addr + (vaddr - (uint64_t)res->addr);
DEBUG_PRINT("%s: %p -> %p\n", __func__, (void *)vaddr,
(void *)paddr);
pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
return paddr;
}
}
}
pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
return SPDK_VTOPHYS_ERROR;
}
static int
spdk_vtophys_notify(void *cb_ctx, struct spdk_mem_map *map,
enum spdk_mem_map_notify_action action,
void *vaddr, size_t len)
{
int rc = 0, pci_phys = 0;
uint64_t paddr;
if ((uintptr_t)vaddr & ~MASK_256TB) {
DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
return -EINVAL;
}
if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
DEBUG_PRINT("invalid parameters, vaddr=%p len=%ju\n",
vaddr, len);
return -EINVAL;
}
/* Get the physical address from the DPDK memsegs */
paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
switch (action) {
case SPDK_MEM_MAP_NOTIFY_REGISTER:
if (paddr == SPDK_VTOPHYS_ERROR) {
/* This is not an address that DPDK is managing. */
#if SPDK_VFIO_ENABLED
enum rte_iova_mode iova_mode;
#if RTE_VERSION >= RTE_VERSION_NUM(19, 11, 0, 0)
iova_mode = rte_eal_iova_mode();
#else
iova_mode = rte_eal_get_configuration()->iova_mode;
#endif
if (spdk_iommu_is_enabled() && iova_mode == RTE_IOVA_VA) {
/* We'll use the virtual address as the iova to match DPDK. */
paddr = (uint64_t)vaddr;
rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, len);
if (rc) {
return -EFAULT;
}
while (len > 0) {
rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
if (rc != 0) {
return rc;
}
vaddr += VALUE_2MB;
paddr += VALUE_2MB;
len -= VALUE_2MB;
}
} else
#endif
{
/* Get the physical address from /proc/self/pagemap. */
paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
if (paddr == SPDK_VTOPHYS_ERROR) {
/* Get the physical address from PCI devices */
paddr = vtophys_get_paddr_pci((uint64_t)vaddr);
if (paddr == SPDK_VTOPHYS_ERROR) {
DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
return -EFAULT;
}
/* The beginning of this address range points to a PCI resource,
* so the rest must point to a PCI resource as well.
*/
pci_phys = 1;
}
/* Get paddr for each 2MB chunk in this address range */
while (len > 0) {
/* Get the physical address from /proc/self/pagemap. */
if (pci_phys) {
paddr = vtophys_get_paddr_pci((uint64_t)vaddr);
} else {
paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
}
if (paddr == SPDK_VTOPHYS_ERROR) {
DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
return -EFAULT;
}
/* Since PCI paddr can break the 2MiB physical alignment skip this check for that. */
if (!pci_phys && (paddr & MASK_2MB)) {
DEBUG_PRINT("invalid paddr 0x%" PRIx64 " - must be 2MB aligned\n", paddr);
return -EINVAL;
}
#if SPDK_VFIO_ENABLED
/* If the IOMMU is on, but DPDK is using iova-mode=pa, we want to register this memory
* with the IOMMU using the physical address to match. */
if (spdk_iommu_is_enabled()) {
rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, VALUE_2MB);
if (rc) {
DEBUG_PRINT("Unable to assign vaddr %p to paddr 0x%" PRIx64 "\n", vaddr, paddr);
return -EFAULT;
}
}
#endif
rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
if (rc != 0) {
return rc;
}
vaddr += VALUE_2MB;
len -= VALUE_2MB;
}
}
} else {
/* This is an address managed by DPDK. Just setup the translations. */
while (len > 0) {
paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
if (paddr == SPDK_VTOPHYS_ERROR) {
DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
return -EFAULT;
}
rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
if (rc != 0) {
return rc;
}
vaddr += VALUE_2MB;
len -= VALUE_2MB;
}
}
break;
case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
#if SPDK_VFIO_ENABLED
if (paddr == SPDK_VTOPHYS_ERROR) {
/*
* This is not an address that DPDK is managing. If vfio is enabled,
* we need to unmap the range from the IOMMU
*/
if (spdk_iommu_is_enabled()) {
uint64_t buffer_len = len;
uint8_t *va = vaddr;
enum rte_iova_mode iova_mode;
#if RTE_VERSION >= RTE_VERSION_NUM(19, 11, 0, 0)
iova_mode = rte_eal_iova_mode();
#else
iova_mode = rte_eal_get_configuration()->iova_mode;
#endif
/*
* In virtual address mode, the region is contiguous and can be done in
* one unmap.
*/
if (iova_mode == RTE_IOVA_VA) {
paddr = spdk_mem_map_translate(map, (uint64_t)va, &buffer_len);
if (buffer_len != len || paddr != (uintptr_t)va) {
DEBUG_PRINT("Unmapping %p with length %lu failed because "
"translation had address 0x%" PRIx64 " and length %lu\n",
va, len, paddr, buffer_len);
return -EINVAL;
}
rc = vtophys_iommu_unmap_dma(paddr, len);
if (rc) {
DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
return -EFAULT;
}
} else if (iova_mode == RTE_IOVA_PA) {
/* Get paddr for each 2MB chunk in this address range */
while (buffer_len > 0) {
paddr = spdk_mem_map_translate(map, (uint64_t)va, NULL);
if (paddr == SPDK_VTOPHYS_ERROR || buffer_len < VALUE_2MB) {
DEBUG_PRINT("could not get phys addr for %p\n", va);
return -EFAULT;
}
rc = vtophys_iommu_unmap_dma(paddr, VALUE_2MB);
if (rc) {
DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
return -EFAULT;
}
va += VALUE_2MB;
buffer_len -= VALUE_2MB;
}
}
}
}
#endif
while (len > 0) {
rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
if (rc != 0) {
return rc;
}
vaddr += VALUE_2MB;
len -= VALUE_2MB;
}
break;
default:
SPDK_UNREACHABLE();
}
return rc;
}
static int
vtophys_check_contiguous_entries(uint64_t paddr1, uint64_t paddr2)
{
/* This function is always called with paddrs for two subsequent
* 2MB chunks in virtual address space, so those chunks will be only
* physically contiguous if the physical addresses are 2MB apart
* from each other as well.
*/
return (paddr2 - paddr1 == VALUE_2MB);
}
#if SPDK_VFIO_ENABLED
static bool
spdk_vfio_enabled(void)
{
return rte_vfio_is_enabled("vfio_pci");
}
/* Check if IOMMU is enabled on the system */
static bool
has_iommu_groups(void)
{
struct dirent *d;
int count = 0;
DIR *dir = opendir("/sys/kernel/iommu_groups");
if (dir == NULL) {
return false;
}
while (count < 3 && (d = readdir(dir)) != NULL) {
count++;
}
closedir(dir);
/* there will always be ./ and ../ entries */
return count > 2;
}
static bool
spdk_vfio_noiommu_enabled(void)
{
return rte_vfio_noiommu_is_enabled();
}
static void
spdk_vtophys_iommu_init(void)
{
char proc_fd_path[PATH_MAX + 1];
char link_path[PATH_MAX + 1];
const char vfio_path[] = "/dev/vfio/vfio";
DIR *dir;
struct dirent *d;
if (!spdk_vfio_enabled()) {
return;
}
if (spdk_vfio_noiommu_enabled()) {
g_vfio.noiommu_enabled = true;
} else if (!has_iommu_groups()) {
return;
}
dir = opendir("/proc/self/fd");
if (!dir) {
DEBUG_PRINT("Failed to open /proc/self/fd (%d)\n", errno);
return;
}
while ((d = readdir(dir)) != NULL) {
if (d->d_type != DT_LNK) {
continue;
}
snprintf(proc_fd_path, sizeof(proc_fd_path), "/proc/self/fd/%s", d->d_name);
if (readlink(proc_fd_path, link_path, sizeof(link_path)) != (sizeof(vfio_path) - 1)) {
continue;
}
if (memcmp(link_path, vfio_path, sizeof(vfio_path) - 1) == 0) {
sscanf(d->d_name, "%d", &g_vfio.fd);
break;
}
}
closedir(dir);
if (g_vfio.fd < 0) {
DEBUG_PRINT("Failed to discover DPDK VFIO container fd.\n");
return;
}
g_vfio.enabled = true;
return;
}
#endif
void
spdk_vtophys_pci_device_added(struct rte_pci_device *pci_device)
{
struct spdk_vtophys_pci_device *vtophys_dev;
pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
vtophys_dev = calloc(1, sizeof(*vtophys_dev));
if (vtophys_dev) {
vtophys_dev->pci_device = pci_device;
TAILQ_INSERT_TAIL(&g_vtophys_pci_devices, vtophys_dev, tailq);
} else {
DEBUG_PRINT("Memory allocation error\n");
}
pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
#if SPDK_VFIO_ENABLED
struct spdk_vfio_dma_map *dma_map;
int ret;
if (!g_vfio.enabled) {
return;
}
pthread_mutex_lock(&g_vfio.mutex);
g_vfio.device_ref++;
if (g_vfio.device_ref > 1) {
pthread_mutex_unlock(&g_vfio.mutex);
return;
}
/* This is the first SPDK device using DPDK vfio. This means that the first
* IOMMU group might have been just been added to the DPDK vfio container.
* From this point it is certain that the memory can be mapped now.
*/
TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
if (ret) {
DEBUG_PRINT("Cannot update DMA mapping, error %d\n", errno);
break;
}
}
pthread_mutex_unlock(&g_vfio.mutex);
#endif
}
void
spdk_vtophys_pci_device_removed(struct rte_pci_device *pci_device)
{
struct spdk_vtophys_pci_device *vtophys_dev;
pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
if (vtophys_dev->pci_device == pci_device) {
TAILQ_REMOVE(&g_vtophys_pci_devices, vtophys_dev, tailq);
free(vtophys_dev);
break;
}
}
pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
#if SPDK_VFIO_ENABLED
struct spdk_vfio_dma_map *dma_map;
int ret;
if (!g_vfio.enabled) {
return;
}
pthread_mutex_lock(&g_vfio.mutex);
assert(g_vfio.device_ref > 0);
g_vfio.device_ref--;
if (g_vfio.device_ref > 0) {
pthread_mutex_unlock(&g_vfio.mutex);
return;
}
/* This is the last SPDK device using DPDK vfio. If DPDK doesn't have
* any additional devices using it's vfio container, all the mappings
* will be automatically removed by the Linux vfio driver. We unmap
* the memory manually to be able to easily re-map it later regardless
* of other, external factors.
*/
TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &dma_map->unmap);
if (ret) {
DEBUG_PRINT("Cannot unmap DMA memory, error %d\n", errno);
break;
}
}
pthread_mutex_unlock(&g_vfio.mutex);
#endif
}
int
spdk_vtophys_init(void)
{
const struct spdk_mem_map_ops vtophys_map_ops = {
.notify_cb = spdk_vtophys_notify,
.are_contiguous = vtophys_check_contiguous_entries,
};
const struct spdk_mem_map_ops phys_ref_map_ops = {
.notify_cb = NULL,
.are_contiguous = NULL,
};
#if SPDK_VFIO_ENABLED
spdk_vtophys_iommu_init();
#endif
g_phys_ref_map = spdk_mem_map_alloc(0, &phys_ref_map_ops, NULL);
if (g_phys_ref_map == NULL) {
DEBUG_PRINT("phys_ref map allocation failed.\n");
return -ENOMEM;
}
g_vtophys_map = spdk_mem_map_alloc(SPDK_VTOPHYS_ERROR, &vtophys_map_ops, NULL);
if (g_vtophys_map == NULL) {
DEBUG_PRINT("vtophys map allocation failed\n");
return -ENOMEM;
}
return 0;
}
uint64_t
spdk_vtophys(void *buf, uint64_t *size)
{
uint64_t vaddr, paddr_2mb;
vaddr = (uint64_t)buf;
paddr_2mb = spdk_mem_map_translate(g_vtophys_map, vaddr, size);
/*
* SPDK_VTOPHYS_ERROR has all bits set, so if the lookup returned SPDK_VTOPHYS_ERROR,
* we will still bitwise-or it with the buf offset below, but the result will still be
* SPDK_VTOPHYS_ERROR. However now that we do + rather than | (due to PCI vtophys being
* unaligned) we must now check the return value before addition.
*/
SPDK_STATIC_ASSERT(SPDK_VTOPHYS_ERROR == UINT64_C(-1), "SPDK_VTOPHYS_ERROR should be all 1s");
if (paddr_2mb == SPDK_VTOPHYS_ERROR) {
return SPDK_VTOPHYS_ERROR;
} else {
return paddr_2mb + (vaddr & MASK_2MB);
}
}